Module: tf.contrib.summary

View source on GitHub

TensorFlow Summary API v2.

The operations in this package are safe to use with eager execution turned on or off. It has a more flexible API that allows summaries to be written directly from ops to places other than event log files, rather than propagating protos from tf.summary.merge_all to tf.summary.FileWriter.

To use with eager execution enabled, write your code as follows:

global_step = tf.train.get_or_create_global_step()
summary_writer = tf.contrib.summary.create_file_writer(
    train_dir, flush_millis=10000)
with summary_writer.as_default(), tf.contrib.summary.always_record_summaries():
  # model code goes here
  # and in it call
  tf.contrib.summary.scalar("loss", my_loss)
  # In this case every call to tf.contrib.summary.scalar will generate a record
  # ...

To use it with graph execution, write your code as follows:

global_step = tf.train.get_or_create_global_step()
summary_writer = tf.contrib.summary.create_file_writer(
    train_dir, flush_millis=10000)
with summary_writer.as_default(), tf.contrib.summary.always_record_summaries():
  # model definition code goes here
  # and in it call
  tf.contrib.summary.scalar("loss", my_loss)
  # In this case every call to tf.contrib.summary.scalar will generate an op,
  # note the need to run tf.contrib.summary.all_summary_ops() to make sure these
  # ops get executed.
  # ...
  train_op = ....

with tf.Session(...) as sess:
  tf.global_variables_initializer().run()
  tf.contrib.summary.initialize(graph=tf.get_default_graph())
  # ...
  while not_done_training:
    sess.run([train_op, tf.contrib.summary.all_summary_ops()])
    # ...

Classes

class SummaryWriter: Interface representing a stateful summary writer object.

Functions

all_summary_ops(...): Returns all V2-style summary ops defined in the current default graph.

always_record_summaries(...): Sets the should_record_summaries Tensor to always true.

audio(...): Writes an audio summary if possible.

create_db_writer(...): Creates a summary database writer in the current context.

create_file_writer(...): Creates a summary file writer in the current context under the given name.

create_summary_file_writer(...): Please use tf.contrib.summary.create_file_writer. (deprecated)

eval_dir(...): Construct a logdir for an eval summary writer.

flush(...): Forces summary writer to send any buffered data to storage.

generic(...): Writes a tensor summary if possible.

graph(...): Writes a TensorFlow graph to the summary interface.

histogram(...): Writes a histogram summary if possible.

image(...): Writes an image summary if possible.

import_event(...): Writes a tf.compat.v1.Event binary proto.

initialize(...): Initializes summary writing for graph execution mode.

never_record_summaries(...): Sets the should_record_summaries Tensor to always false.

record_summaries_every_n_global_steps(...): Sets the should_record_summaries Tensor to true if global_step % n == 0.

scalar(...): Writes a scalar summary if possible.

should_record_summaries(...): Returns boolean Tensor which is true if summaries should be recorded.

summary_writer_initializer_op(...): Graph-mode only. Returns the list of ops to create all summary writers.